Lishuai Jing

Grep Is Not a Memory

High-fidelity context for coding agents: structure, time, and live updates

Every coding agent understands your codebase the same way. It greps, opens files, pieces together how things connect, acts on that guess, and forgets all of it when the session ends. Tomorrow it starts again from zero and you pay for the same reading twice. It can't tell which of three functions called verify is the one that matters, doesn't know that a signature changed an hour ago, has no idea why that odd retry wrapper exists, and can't see the agent working next to it. So it confidently "fixes" working code, and you spend the afternoon cursing at it.

Bigger context windows add more haystack. Better harnesses add rules, but a rule is useless without the facts to apply it to. This talk makes the case for a high-fidelity context layer: a deterministic graph of your code, parsed rather than guessed, aware of time, updated on every save, running locally and served over MCP. And it shows what that changes in practice: better code and reviews, safer refactors, documentation that stays true, and less time and fewer tokens per task.

What attendees take away

1. Why "more context" and "a stronger harness" don't solve context quality, and exactly where each one breaks.
2. Four properties to evaluate any context layer: structural, temporal, live, local.
3. What high-fidelity context changes day to day across quality, review, refactoring, docs and cost, and what to measure to prove it on your own team.

Bio:
Lishuai Jing is co-founder and CAIO of Memtrace, a Copenhagen-based AI infrastructure startup backed by Antler and Innovation Fund Denmark. Memtrace gives AI coding agents a deterministic, knowledge-graph memory of the codebase, built through AST parsing rather than guesswork. Before founding Memtrace, Lishuai spent a decade in enterprise AI as Chief Data Scientist at Ønskesken, Lead data scientist at VELUX, and senior data scientist at Grundfos.